Children use statistical learning in naturalistic language acquisition: evidence from a large longitudinal corpus of a polysynthetic language

Stoll, S. . 1 , Mazara, J. 1 , Blasi, D. 1, 2 & Bickel, B. 1

1 Department of Comparative Linguistics, University of Zurich, Switzerland
2 Max Planck Institute for the Science of Human History, Jena, Germany

Experiments with controlled frequency distributions suggest that statistical learning is a powerful learning mechanism. Here, we move beyond the proof of concept stage in showing that children use statistical learning in naturalistic acquisition. Evidence comes from a high-density longitudinal study of 4 Chintang children (Sino-Tibetan, Nepal) age 2;0-4;6. Chintang has an extremely complex verbal system with nearly 5000 verb forms. Many of them are produced rarely resulting in skewed input distributions, which challenge the universal status of statistical learning.
We show that in verb learning Chintang children rely on sublexical regularities within verbs rather than on full form regularities. Chintang verbs consist of two parts: heads (stems), which are frequently repeated and dependents (grammatical affixes), which are less frequently repeated. A comparison of verb entropies by adults and children shows that the use of heads approaches adult performance earlier than the use of dependents. Around age 4 children are similar in their overall distributions to adults.
Our results suggest that the acquisition process successfully adapts to even the most complex languages because the statistical learning mechanism flexibly focuses on language-specific regularities and detects patterns provided in the input.